Deep Learning based drift field correction for MR Thermometry in the upper leg at 7T

PRFS-based MR Thermometry (MRT) bears strong potential for RF safety assessment. However, PRFS-MRT is impaired by external sources of frequency shift. It is hypothesized that deep learning will be able to separate the PRFS signal from these other sources of frequency shift. This study has tested this concept on drift field correction for MRT in the human thigh at 7T. A convolutional neural network is trained using synthetic phase difference images based on measured drift fields and simulated temperature distributions. Results show that the proposed deep-learning approach is able to correctly predict both simulated and measured temperature rise distributions.